1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/neon-p5.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x4(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x4(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n % sizeof(float) == 0);
25
26 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->neonfma_rr1_p5.magic_bias);
27 const float32x4_t vminus_log2e = vld1q_dup_f32(¶ms->neonfma_rr1_p5.minus_log2e);
28 const float32x4_t vln2 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.ln2);
29 const float32x4_t vc5 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c5);
30 const float32x4_t vc4 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c4);
31 const float32x4_t vc3 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c3);
32 const float32x4_t vc2 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c2);
33 const float32x4_t vc1 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c1);
34 const float32x4_t vone = vmovq_n_f32(1.0f);
35 const float32x4_t vdenorm_cutoff = vld1q_dup_f32(¶ms->neonfma_rr1_p5.denorm_cutoff);
36
37 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
38 const float32x4_t vx = vld1q_f32(x); x += 4;
39
40 const float32x4_t vz = vabsq_f32(vx);
41
42 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
43 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
44 vn = vsubq_f32(vn, vmagic_bias);
45 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
46
47 float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
48 vp = vfmaq_f32(vc3, vp, vt);
49 vp = vfmaq_f32(vc2, vp, vt);
50 vp = vfmaq_f32(vc1, vp, vt);
51
52 vt = vmulq_f32(vt, vs);
53 const float32x4_t ve = vfmaq_f32(vs, vp, vt);
54 const float32x4_t vd = vaddq_f32(ve, vone);
55
56 float32x4_t vr = vrecpeq_f32(vd);
57 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
58 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
59
60 float32x4_t vf = vmulq_f32(ve, vr);
61 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
62 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
63 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
64
65 vst1q_f32(y, vf); y += 4;
66 }
67 if XNN_UNLIKELY(n != 0) {
68 const float32x4_t vx = vld1q_f32(x);
69
70 const float32x4_t vz = vabsq_f32(vx);
71
72 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
73 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
74 vn = vsubq_f32(vn, vmagic_bias);
75 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
76
77 float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
78 vp = vfmaq_f32(vc3, vp, vt);
79 vp = vfmaq_f32(vc2, vp, vt);
80 vp = vfmaq_f32(vc1, vp, vt);
81
82 vt = vmulq_f32(vt, vs);
83 const float32x4_t ve = vfmaq_f32(vs, vp, vt);
84 const float32x4_t vd = vaddq_f32(ve, vone);
85
86 float32x4_t vr = vrecpeq_f32(vd);
87 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
88 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
89
90 float32x4_t vf = vmulq_f32(ve, vr);
91 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
92 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
93 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
94
95 float32x2_t vf_lo = vget_low_f32(vf);
96 if (n & (2 * sizeof(float))) {
97 vst1_f32(y, vf_lo); y += 2;
98 vf_lo = vget_high_f32(vf);
99 }
100 if (n & (1 * sizeof(float))) {
101 vst1_lane_f32(y, vf_lo, 0);
102 }
103 }
104 }
105